Blog Engineering How GitLab can help mitigate deletion of open source container images on Docker Hub
Published on: March 16, 2023
13 min read

How GitLab can help mitigate deletion of open source container images on Docker Hub

CI/CD and Kubernetes deployments can be affected by Docker Hub tier changes. This tutorial walks through analysis, mitigations, and long-term solutions.


Docker, Inc. shared an email update to Docker Hub users that it will sunset Free Team organizations. If accounts do not upgrade to a paid plan before April 14, 2023, their organization's images may be deleted after 30 days. This change can affect open source organizations that publish their images on Docker Hub, as well as consumers of these container images, used in CI/CD pipelines, Kubernetes cluster deployments, or docker-compose demo environments. This blog post discusses tools and features on the GitLab DevSecOps platform to help users analyze and mitigate the potential impact on production environments.

Update (March 20, 2023): Docker, Inc. published an apology blog post, including a FAQ, and clarifies that the company will not delete container images by themselves. Maintainers can migrate to a personal account, join the Docker-sponsored open source program, or opt into a paid plan. If open source container image maintainers do nothing, this leads into another issue: Stale container images can become a security problem. The following blog post can help with security analysis and migration too.

Update (March 27, 2023): On March 24, 2023, Docker, Inc. published another blog post announcing the reversal of the decision to sunset the Free team plan and updated its FAQ for Free Team organization. While this is a welcome development for the entire community, it is still crucial to ensure the reliability of your software development lifecycle by ensuring redundancies are in place for your container registries, as detailed in this blog post.

Inventory of used container images

CI/CD pipelines in GitLab can execute jobs in containers. This is specified by the image keyword in jobs, job templates, or as a global default attribute. For the first iteration, you can clone a GitLab project locally, and search for the image string in all CI/CD configuration files. The following example shows how to execute the find command on the command line interface (CLI), searching for files matching the name pattern *ci.yml, and looking for the image string in the file content. The command line prints a list of search pattern matches, and the corresponding file name to the standard output. The example inspects the project for the GitLab handbook and website to analyze whether its CI/CD deployment pipelines could be affected by the Docker Hub changes.

$ git clone && cd www-gitlab-com

$ find . -type f -iname '*ci.yml' -exec sh -c "grep 'image:' '{}' && echo {}" \;

  image: alpine:edge
  image: alpine:edge
  image: debian:stable-slim
  image: debian:stable-slim

A discussion on Hacker News mentions that "official Docker images" are not affected, but this is not officially confirmed by Docker yet. Official Docker images do not use a namespace prefix, i.e. namespace/imagename but instead debian:<tagname> for example. uses a full URL image string, which includes the image registry server domain, in the shown example.

If there is no full URL prefix in the image string, this is an indicator that this image could be pulled from Docker Hub by default. There might be other infrastructure safety nets put in place, for example a cloud provider registry which caches the Docker Hub images (Google Cloud, AWS, Azure, etc.).

Advanced search for images

You can use the project lint API endpoint to fetch the CI configuration. The following script uses the python-gitlab API library to implement the API endpoint:

  1. Collect all projects from either a single project ID, a group ID with projects, or from the instance.
  2. Run the project.ci_lint.get() method to get a merged yaml configuration for CI/CD from the current GitLab project.
  3. Parse the yaml content and print only the job names, and the image keys.

The full script is located here, and is open source, licensed under MIT.

#!/usr/bin/env python

import gitlab
import os
import sys
import yaml

GITLAB_SERVER = os.environ.get('GL_SERVER', '')
GITLAB_TOKEN = os.environ.get('GL_TOKEN') # token requires developer permissions
PROJECT_ID = os.environ.get('GL_PROJECT_ID') #optional
GROUP_ID = os.environ.get('GL_GROUP_ID', 65096153) #optional

# Main

if __name__ == "__main__":
    if not GITLAB_TOKEN:
        print("🤔 Please set the GL_TOKEN env variable.")

    gl = gitlab.Gitlab(GITLAB_SERVER, private_token=GITLAB_TOKEN)

    # Collect all projects, or prefer projects from a group id, or a project id
    projects = []

    # Direct project ID
    if PROJECT_ID:

    # Groups and projects inside
    elif GROUP_ID:
        group = gl.groups.get(GROUP_ID)

        for project in group.projects.list(include_subgroups=True, all=True):
            manageable_project = gl.projects.get(

    # All projects on the instance (may take a while to process)
        projects = gl.projects.list(get_all=True)

    print("# Summary of projects and their CI/CD image usage")

    # Loop over projects, fetch .gitlab-ci.yml, run the linter to get the full translated config, and extract the `image:` setting
    for project in projects:

        print("# Project: {name}, ID: {id}\n\n".format(name=project.name_with_namespace,

        lint_result = project.ci_lint.get()

        data = yaml.safe_load(lint_result.merged_yaml)

        for d in data:
            print("Job name: {n}".format(n=d))
            for attr in data[d]:
                if 'image' in attr:
                    print("Image: {i}".format(i=data[d][attr]))



The script requires Python (tested with 3.11) and the python-gitlab and pyyaml modules. Example on macOS with Homebrew:

$ brew install python
$ pip3 install python-gitlab pyyaml

You can execute the script and set the different environment variables to control its behavior:


$ export GL_GROUP_ID=12345
$ export GL_PROJECT_ID=98765

$ python3

# Summary of projects and their CI/CD image usage
# Project: Developer Evangelism at GitLab  / use-cases / Docker Use cases  / Custom Container Image Python, ID: 44352983

Job name: docker-build
Image: docker:latest

# Project: Developer Evangelism at GitLab  / use-cases / Docker Use cases  / Gitlab Dependency Proxy, ID: 44351128

Job name: .test-python-version
Job name: image-docker-hub
Image: python:3.11
Job name: image-docker-hub-dep-proxy

Please verify the script and fork it for your own analysis and mitigation. The missing parts are checking the image URLs, and doing a more sophisticated search. The code has been prepared to either check against a single project, a group with projects, or an instance (this may take very long, use with care).

You can perform a more history-focused analysis by fetching the CI/CD job logs from GitLab and search for the pulled container image to get an overview of past Docker executor runs – for example: Using Docker executor with image python:3.11 .... The screenshot shows the CI/CD job logs UI search – you can automate the search using the GitLab API, and the python-gitlab library, for example.

GitLab CI/CD job logs, searching for the  keyword

This snippet can be used in combination with the code shared for the CI lint API endpoint. It fetches the job trace logs, and searches for the image keyword in the log. The missing parts are splitting the log line by line, and extracting the image key information. This is left as an exercise for the reader.

        for job in
            log_trace = str(job.trace())


            if 'image' in log_trace:
                print("Job ID: {i}, URL {u}".format(, u=job.web_url))

More inventory considerations

Similar to the API script for CI/CD navigating through all projects, you will need to analyze all Kubernetes manifest configuration files – using either a pull- or push-based approach. This can be achieved by using the python-gitlab methods to load files from the repository and searching the content in similar ways. Helm charts use container images, too, and will require additional analysis.

An additional search possibility: Custom-built container images that use Docker Hub images as a source. A project will consist of:

  1. Dockerfile file that uses FROM <imagename>
  2. .gitlab-ci.yml configuration file that builds container images (using Docker-in-Docker, Kaniko, etc.)

An alternative search method for customers is available by using the Advanced Search through the GitLab UI and API. The following example uses the scope: blobs to search for the FROM string:

$ export GITLAB_TOKEN=xxxxxxxxx

# Search in

$ curl --header "PRIVATE-TOKEN: $GITLAB_TOKEN" "*"

Command line output from Advanced Search API, scope blobs, search  in  file names.

Mitigations and solutions

The following sections discuss potential mitigation strategies, and long-term solutions.

Mitigation: GitLab dependency proxy

The dependency proxy provides a caching mechanism for Docker Hub images. It helps reduce the bandwidth and time required to download and pull the images. It also helped to mitigate the Docker Hub pull rate limits introduced in 2020. The dependency proxy can be configured for public and private projects.

The dependency proxy needs to be enabled for a group. It also needs to be enabled by an instance administrator for self-managed environments, if turned off.

The following example creates two jobs: image-docker-hub and image-docker-hub-dep-proxy. The dependency proxy job uses the CI_DEPENDENCY_PROXY_GROUP_IMAGE_PREFIX CI/CD variable to instruct GitLab to store the image in the cache, and only pull it once when not available.

    - echo "Testing Python version:"
    - python --version

  extends: .test-python-version
  image: python:3.11

  extends: .test-python-version

The configuration is available in this project.

The stored container image is visible at the group level in the Package and container registries > Dependency Proxy menu.

Mitigation: Container registry mirror

This blog post describes how to run a local container registry mirror. Skopeo from Red Hat is another alternative for syncing container image registries, a practical example is described in this article.

The GitLab Cloud Native installation (Helm charts and Operator) use a mirror of tagged images consumed by the related projects. Other product stages follow a similar approach, the security scanners are shipped in container images maintained by GitLab. This also enables self-managed airgapped installations.

Mitigation: Custom images in GitLab container registry

Reproducible builds and compliance requirements may have required you to create custom container images for CI/CD and Kubernetes already. This is also key to verify that no untested and untrusted images are being used in production. GitLab provides a fully integrated container registry, which can be used natively within CI/CD pipelines and GitOps workflows with the agent for Kubernetes.

The following Dockerfile example extends an existing image layer, and installs additional tools using the Debian Apt package manager.

FROM python:3.11-bullseye

ENV DEBIAN_FRONTEND noninteractive

RUN apt update && apt -y install git curl jq && rm -rf /var/lib/apt/lists/*

You can use Docker to build container images, and alternative options are Kaniko or Podman. On SaaS, you can use the Docker CI/CD template to build and push images. The following example modifies the docker-build job to only build the latest tag from the default branch:

  - template: Docker.gitlab-ci.yml

  stage: build
      #when: manual
      #allow_failure: true

For this example, we specifically want to provide a Git tag that gets used for the container image tag as well.

$ git tag 3-11-bullseye
$ git push --tags

The image will be available at the GitLab container registry URL and the project namespace path.This path needs to be replaced in all projects that use a Python-based image. You can create scripts for the GitLab API to update files and create MRs automatically,


Note: This is a demo project and not actively maintained. Please fork/copy it for your own needs.

Observability and security

The number of failed CI/CD pipelines can be a good service level indicator (SLI) to verify whether the environment is affected by the Docker Hub changes. The same SLI applies for CI/CD jobs that build container images, using a Dockerfile file, which is based on Docker Hub images (FROM ).

A similar SLI applies to Kubernetes cluster deployments – if they continue to generate failures in GitOps pull or CI/CD push scenarios, additional analysis and actions are required. The pod status ErrImagePull and ImagePullBackOff will immediately show the problems. The image pull policy should also be revised – Always will immediately cause a problem, while IfNotPresent will use the local image cache.

This alert rule example for Prometheus observing a Kubernetes cluster can help detect the pod state as not healthy.

  - alert: KubernetesPodNotHealthy
    expr: sum by (namespace, pod) (kube_pod_status_phase{phase=~"Pending|Unknown|Failed"}) > 0
    for: 15m
      severity: critical
      summary: Kubernetes Pod not healthy (instance {{ $labels.instance }})
      description: "Pod has been in a non-ready state for longer than 15 minutes.\n  VALUE = {{ $value }}\n  LABELS = {{ $labels }}"

CI/CD pipeline linters and Git hooks can also be helpful to enforce using a GitLab registry URL prefix in all image tags, when new updates to CI/CD configurations are being pushed into merge requests.

Kubernetes deployment images can be controlled through additional integrations with the Open Policy Agent Gatekeeper or Kyverno. Kyverno also allows you to mutate the image registry location, and redirect the pod image to trusted sources.

Operational container scanning in Kubernetes clusters and container scanning in CI/CD pipelines are recommended. This ensures that all images do not expose security vulnerabilities.

Long-term solutions

As a long-term solution, analyze the affected Docker Hub organizations images and match them against your image usage inventory. Some organizations have raised their concerns in this Docker Hub feedback issue. Be sure to identify critical production CI/CD workflows and replace all external dependencies with local maintained images.

Fork/copy project Dockerfile files from the upstream Git repositories, and use them as the single source of truth for custom container builds. This will also require training and documentation for DevSecOps teams, for example optimizing container images for efficient CI/CD pipelines. More DevSecOps efficiency tips can be found in my Chemnitz Linux Days talk about "Efficient DevSecOps Pipelines in a Cloud Native World" (slides).

Please share your ideas and thoughts about Docker Hub change mitigations and tools on the GitLab community forum. Thank you!

Cover image by Roger Hoyles on Unsplash

We want to hear from you

Enjoyed reading this blog post or have questions or feedback? Share your thoughts by creating a new topic in the GitLab community forum. Share your feedback

Ready to get started?

See what your team could do with a unified DevSecOps Platform.

Get free trial

Find out which plan works best for your team

Learn about pricing

Learn about what GitLab can do for your team

Talk to an expert